The purpose of this weekly report is to present the learning outcomes of week 29, with a focus on the improved version of YOLO v2, the pioneering single-stage algorithm in the field of object detection based on deep learning.
Authors Joseph Redmon and Ali Farhadi made significant improvements based on YOLOv1, proposing YOLOv2 and YOLO9000, with a focus on addressing the shortcomings in YOLOv1 recall and localization accuracy. The main content includes batch normalization, high-resolution classifier Anchor box、 The direct position prediction and fine-grained feature relic multi-scale training have upgraded YOLO v1 from multiple perspectives and improved its poor performance in detecting small targets.
This article will provide a detailed explanation of the learning content and summarize the learning content for this week in the final section. This weekly report aims to effectively combine theoretical knowledge with practical applications in this way, providing a summary of the basic content and direction of deep learning learning learning.
1. 概述